Create an Experiment in Azure Machine Learning

1. Open Machine Learning Studio. Use the Preview Environment to become familiar with the new look and feel. Note: Create the workspace in the same region as your lab provided resource group.
2. Create a Compute instance.
   • The Compute must be uniquely named. We can use the name of the Machine Learning Workspace as part of the Compute name to guarantee uniqueness.
   • This will take a few minutes to spin up. Grab a coffee or tea while waiting.
3. Clone the Jupyter Notebook from GitHub.
   • Go to the Notebooks section of Machine Learning Studio.
   • Create a new Python notebook and edit it in Jupyter.
   • In the first cell, run the following:

   !git clone https://github.com/linuxacademy/content-dp100
   

   • After the repo is cloned, close this notebook and open the content-dp100/notebooks/MNIST_AzureExperiment.ipynb notebook, choosing to edit it in Jupyter.

1. Follow the steps in the notebook.
   • Select Python 3.8 - AzureML in the top right corner.
   • Run each code cell in the MNIST_AzureExperiment notebook and view the results.
   • Read the explanations to get a better understanding of what each cell is doing.

1. View the results in Machine Learning Studio.
   • The final cell of the notebook provides a link directly to the mnist experiment. Run the cell and click the link.
   • Change the graph showing batch_size to instead show accuracy.
   • Remove the Compute target, Job type, and Created by columns from the table.
   • Add the accuracy column to the table.
2. Evaluate how the changing batch_size affects the loss, accuracy, and training time. What would explain this phenomenon?
3. Do Further Research.
   • Local Minima and Maxima
   • Stochastic Gradient Descent

The larger batch sizes are prone to becoming stuck in local minima or missing minima entirely. The models train much faster because less backpropagation has to happen in the network, but the cost is accuracy. Batch sizes between 32 and 256 are good starting points for most models, but this requires tuning like all hyperparameters.